Sentiment Composition
Karo Moilanen and Stephen Pulman
Oxford University Computing Laboratory
Wolfson Building, Parks Road, Oxford, OX1 3QD, England
{ Karo.Moilanen | Stephen.Pulman }@comlab.ox.ac.uk
Abstract
Sentiment classification of grammatical constituents can be explained in a quasicompositional way. The classification of a
complex constituent is derived via the classification of its component constituents and
operations on these that resemble the usual
methods of compositional semantic analysis.
This claim is illustrated with a description of
sentiment propagation, polarity reversal, and
polarity conflict resolution within various linguistic constituent types at various grammatical levels. We propose a theoretical composition model, evaluate a lexical dependency
parsing post-process implementation, and estimate its impact on general NLP pipelines.
Keywords
sentence-level sentiment, clause-level sentiment, entity-level sentiment, valence shifters, polarity shifters, lexical semantics
1
Introduction
Using lists of positive and negative keywords can give
the beginnings of a sentiment classification system.
However, classifying sentiment on the basis of individual words can give misleading results because
atomic sentiment carriers can be modified (weakened,
strengthened, or reversed) based on lexical, discoursal, or paralinguistic contextual operators ([7]). Past
attempts to deal with this phenomenon include writing heuristic rules to look out for negatives and other
‘changing’ words ([6]), combining the scores of individual positive and negative word frequencies ([11], [5]),
and training a classifier on a set of contextual features
([10]). While statistical sentiment classifiers work well
with a sufficiently large input (e.g. a 750-word movie
review), smaller subsentential text units such as individual clauses or noun phrases pose a challenge. It
is such low-level units that are needed for accurate
entity-level sentiment analysis to assign (local) polarities to individual mentions of people, for example.
In this paper we argue that, as far as low-level
(sub)sentential sentiment classification is concerned,
there may be much to be gained from taking account
of more linguistic structure than is usually the case.
In particular we argue that it is possible to calculate
in a systematic way the polarity values of larger syntactic constituents as some function of the polarities
of their subconstituents, in a way almost exactly analogous to the ‘principle of compositionality’ familiar
from the formal semantics literature ([2]). For if the
meaning of a sentence is a function of the meanings
of its parts then the global polarity of a sentence is
a function of the polarities of its parts. For example,
production rules such as [VPα → Vα + NP] and [Sβ
→ NP + VPβ ] operating on a structure like “America invaded Iraq” would treat the verb “invade” as a
function from the NP meaning to the VP meaning
(i.e. as combining semantically with its direct object
to form a VP). The VP meaning is correspondingly
a function from the NP meaning to the S meaning
(i.e. as combining with a subject to form a sentence).
Analogously, a ‘DECREASE’ verb like “reduce” (cf. [1])
should then be analysed as having a compositional sentiment property such that it reverses the polarity (α)
(¬α)
of its object NP in forming the VP, hence [VPβ
→
(α)
Vβ[DECREASE] + NP ]. Thus the positive polarity
in “reduce the risk” even though “risk” is negative in
itself (cf. the negative polarity in “reduce productivity”). In fact, this semi-compositionality also holds at
other linguistic levels: certainly amongst morphemes,
and arguably also at suprasentential levels. However,
this paper discusses only sentential sentiment composition. Grounded on the descriptive grammatical
framework by ([4]), we propose a theoretical framework within which the sentiment of such structures
can be calculated.
2
Composition Model
The proposed sentiment composition model combines
two input (IN) constituents at a time and calculates
a global polarity for the resultant composite output
(OUT) constituent (cf. parent node dominance in
the modifies polarity and modified by polarity structural features in ([10])). The two IN constituents can
be of any syntactic type or size. The model assumes
dominance of non-neutral (positive (+), negative (-),
mixed (M)) sentiment polarity over neutral (N) polarity. The term sentiment propagation is used here to
denote compositions in which the polarity of a neutral
constituent is overridden by that of a non-neutral constituent ({(+)(N)} → (+); {(-)(N)} → (-)). We use
the term polarity reversal to denote compositions
in which a non-neutral polarity value is changed to
another non-neutral polarity value ((+) → (-); (-) →
(+)) (cf. [7]), and the term polarity conflict to denote compositions containing conflicting non-neutral
polarities ({(+)(-)} → (M)). Polarity conflict resolution refers to disambiguating compositions involving
a polarity conflict ((M) → (+); (M) → (-)).
Polarity conflict resolution is achieved by ranking
the IN constituents on the basis of relative weights
assigned to them dictating which constituent is more
important with respect to sentiment. The stronger of
the IN constituents is here denoted as SPR (superordinate) whereas the label SUB (subordinate) refers
to the dominated constituent (i.e. SPR SUB). Except for (N)[=] SPR constituents, it is therefore the
SPR constituent and the compositional processes executed by it that determine the polarity (α) of the
OUT constituent (i.e. OUTαij → SPRαi + SUBαj ).
The weights are not properties of individual IN constituents per se but are latent in specific syntactic constructions such as [Mod:Adj Head:N] (e.g. adjectival
premodification of head nouns) or [Head:V Comp:NP]
(e.g. direct object complements of verbs).
We tag each entry in the sentiment lexica (across
all word classes) and each constituent with one of the
following tags: default ([=]), positive ([+]), negative ([-]), and reverse ([¬]). These tags allow us to
specify at any structural level and composition stage
what any given SPR constituent does locally to the
polarity of an accompanying SUB constituent without fixed-order windows of n tokens (cf. ([7]), modification features in ([10]), change phrases in ([6])).
A [=] SPR constituent combines with a SUB constituent in the default fashion. The majority of constituents are [=]. A [¬] SPR constituent reverses the
polarity of the SUB constituent and assigns that polarity to the OUT constituent (cf. general polarity
shifters in ([10])). As SPR constituents, some carriers
such as “[contaminate](-) ” or “[soothe](+) ” exhibit
such strong sentiment that they can determine the
OUT polarity irrespective of the SUB polarity - consider the static negativity in “[contaminated that damn
disk](-) ”, “[contaminated the environment](-) ”, and
“[contaminated our precious water](-) ” (vice versa for
some positive carriers). Hence the [-] and [+] constants
which can furthermore be used as polarity heuristics
for carriers occurring prototypically with a specific polarity (e.g. “[deficiency (of sth positive)](-) ”) (cf. presuppositional items in ([7]), negative and positive polarity shifters in ([10])).
Notice that the SPR constituent operates on the
SUB constituent irrespective of the polarity of the
latter as a [¬] SPR constituent such as the determiner “[less](N)[¬] ” reverses both (+) and (-) SUB
constituents (e.g. “[less tidy](-) ”, “[less ugly](+) ”),
for example. However, cases in which SPR operations are required only in conjunction with a specific
SUB constituent polarity do exist. The reversal potential in the degree modifier “[too](N)[¬] ”, for instance,
seems to operate only alongside (+) SUB constituents
(i.e. “[too colourful](-) ” vs. “??[too sad](+) ”). The
adjective “[effective](+)[=] ” operates similarly only
with (+) or (N) SUB constituents (i.e. “[effective
remedies/diagrams](+) ” vs. “[effective torture](-) ”).
It is thus proposed that (?:+) and (?:-) be used as
further filters to block specific SPR polarities as required by individual carriers.
To illustrate how the composition model operates,
consider the sample sentence in Ex. 1:
1) The senators supporting(+) the leader(+)
failed(-) to praise(+) his hopeless(-) HIV(-)
prevention program.
Raw frequency counts, yielding three (+) and three
(-) carriers, would fail to predict the global negative
polarity in the sentence. We represent the sentence
as follows, starting with the direct object NP of the
predicator “[praise](+)[+] ” (Ex. 2):
2)
NP
(-)[=]
Head:Nom
Subj-Det:NPgen
(-)[=]
Head:N
his
Mod:Adj
Head:Nom
(N)[=]
(+)[=]
hopeless
(-)[=]
Mod:Nom
Head:N
(+)[=]
program
Mod:N Head:N
3)
(N)[=]
NP
(+)[=]
HIV prevention
(-)[=]
Det:Det
(N)[¬]
Head:Nom
(+)[=]
The
(N)[=]
Head:N
Comp:VP
(+)[=]
senators
(N)[=]
Head:V
Comp:NP
(+)[=]
supporting
(+)[=]
Det:Det
4)
Head:N
S
(-)[=]
the
leader
(N)[=]
(+)[=]
Comp:NP
Head:VP
(+)[=]
(-)[N]
The senators
Head:V
Comp:VP
(+)[=]
supfailed
porting
(-)[¬]
the
Head:VGrp
Comp:NP
leader
(+)[+]
(-)[=]
Mod:TO Head:V his hopeless
HIV prevento
praise tion program
(N)[=]
(+)[+]
Through polarity reversal, the internal sentiment
in “[HIV prevention](+)[=] ” is first arrived at
due to the [¬] status of the SPR head noun
“[prevention](N)[¬] ” which reverses the (-) premodifying noun “[HIV](-)[=] ”. The (N) head noun
“[program](N)[=] ” is then overridden by the (+) premodifying nominal “[HIV prevention](+)[=] ”. When
the resultant nominal is combined with the premodifying attributive SPR input “[hopeless](-)[=] ”, the ensuing polarity conflict can be resolved through the
dominance of the premodifier in this syntactic situation. The final combination with the SUB subject determiner “[his](N)[=] ” is a case of propagation as the resultant NP reflects the polarity of the
head nominal. Sentiment propagation can be seen
throughout the subject NP (Ex. 3) as the (+) head
noun “[leader](+)[=] ”, combined with a (N) SPR determiner, results in a (+) NP (“[the leader](+)[=] ”).
When that NP is combined with a (+) SPR head participial, a (+) SPR VP is generated (“[supporting the
leader](+)[=] ”) which in turn overrides the (N) head
noun “[senators](N)[=] ”. The final (N) SPR determiner does not change the polarity any further.
The NPs thus resolved can then be combined with
the two predicators to form a sentence (Ex. 4).
The direct object NP “[his hopeless HIV prevention
program](-)[=] ” is reversed when it is combined with an
SPR verb group outputting constant positivity (“[to
praise](+)[+] ”). When the resultant (+) VP is used as
the complement of a [¬] SPR head verb polarity reversal occurs once again yielding a (-) VP (“[failed to
praise his hopeless HIV prevention program](-)[=] ”).
Lastly, the (+) subject NP combines with the (-) predicate, and the polarity conflict is resolved due to the
predicate being the SPR constituent. Hence the global
negative sentiment for the present sample sentence can
be calculated from its subconstituents.
3
Grammatical Constructions
Within a syntactic phrase, the polarity of the phrasal
head can be changed by its pre- and post-modifying
dependents. In general, pre-head dependents dominate their heads. Determiners (e.g. “[no crime](-) ”)
and DPs (e.g. “[too much wealth](-) ”) can be modelled as [Det:(Det|DP) Head:N] ([4]: 354-99, 4312, 549, 573). Attributive pre-head AdjPs and simple pre-head ING/EN Participials are ranked similarly as [Mod:(AdjP|V) Head:N] to account for
polarity reversals (e.g. “[trivial problem](+) ”), conflicts (e.g. “[nasty smile](-) ”), and seemingly contradictory compositions with (?:-) premodifiers (e.g.
“[perfected torture](-) ”). However, mixed sentiment is
possible in this construction (e.g. “[savvy liar](M) ”)
([4]: 444). We rank attributive pre-head Adverbs
as [Mod:Adv Head:(Adj|Adv)] (e.g. “[decreasingly happy](-) ”, “[never graceful(ly)](-) ”) although
they too can lead to unresolvable mixed sentiment
(e.g. “[impressively bad(ly)](M) ”) (idem. 548, 5723, 582-5). The pre-head Negator (Neg) “not”,
which is stronger than its head in NPs (e.g. “[not
a scar](+) ”), AdjPs, AdvPs, and PPs, is ranked
as [Mod:Neg Head:(N|Adj|Adv|P)] (cf. [7]). In
contrast, pre-head Nouns and Nominals in NPs
are secondary ([Head:N Mod:(N|Nom)]) as seen
in polarity conflicts (e.g. “[family benefit fraud](-) ”,
“[abuse helpline](+) ”) and [¬] head nouns (e.g. “[risk
minimisation](+) ”) (idem. 444, 448-9). The genitive
subject determiner with the clitic ’s appears similarly
weaker than its head noun or nominal ([Head:(N|Nom)
Subj-Det:NPgen ]) (e.g. “[the war’s end](+) ”), although polarity conflicts can lead to exceptions: com-
pare “[the offender’s apology](+) ” with “[the rapist’s
smile](-) ” (idem. 467-83).
Post-head dependents’ weights are more variable.
In NPs, post-head AdjPs generally dominate (e.g.
“[my best friend angry at me](-) ”) as [Comp:AdjP Head:N] (idem. 445). Post-head Participials dominate their head nouns as [Comp:VP Head:N] (e.g.
“[ugly kids smiling](+) ”, “[the cysts removed](+) ”)
(idem. 446), but post-head VPs are dominated by
their head prepositions ([Head:P Comp:VP]) (e.g.
“[against helping her](-) ”) ([4]: 641). Post-head PPs
are likewise dominated by their noun, adjective, or
adverb heads. The rankings [Head:(N|Adj|Adv) Comp:PP] are thus proposed (e.g. “[different(ly) from
those losers](+) ”, “[unhappy with success](-) ”, “[the
end of the war](+) ”) ([4]: 446, 543-6). However, exceptions may surface in these constructions, especially in
NPs: compare “[two morons amongst my friends](-) ”
with “[cute kittens near a vicious python](-) ”. Moreover, mixed sentiment may surface (e.g. “[angry
protesters against the war](M) ”). Lastly, we rank
post-head NPs in PPs as [Head:P Comp:NP] (e.g.
“[against racism](+) ”, “[with pleasure](+) ”) (idem.
635).
In clausal analysis, we treat as the clausal head the
predicator (P) which is made of one verb group and
compulsory (C)omplements and optional (A)djuncts.
The predicator is generally stronger than its complements. We propose that internal complements (Direct Object (OD ), Indirect Object (OI ), Subject Predicative Complement (PCS ), Object Predicative Complement (PCO ), and Oblique (C)omplement) be combined with the predicator before combining the resultant predicate with the predicator’s external complements ([4]: 215-8; 236-57). In Monotransitive Predicates (P-OD ), the ranking [Head:P Comp:OD ] models propagation (e.g. “[failed it](-) ”),
polarity conflicts (e.g. “[spoiled the party](-) ”), and
[¬] predicators (e.g. “[prevent the war](+) ”) (idem.
244-8). Ditransitive Predicates (P-OI -OD ), (POD -C) behave in a similar way. Since the monotransitive “[sent junk](-) ”, pure ditransitive “[sent
me junk](-) ”, and oblique ditransitive “[sent junk to
me](-) ” all share a [-] P-OD core, we resolve it first
before adding an OI or C to model propagation (e.g.
“[baked a yummy cake for me](+) ”), and polarity conflicts (e.g. “[brought my friend sad news](-) ”) (idem.
244-8). Through the ranking [Head:P Comp:PCS ],
typically (N) copular verbs in Complex Intransitive Predicates (P-PCS ) can be explained (e.g.
“[seems nice](+) ”) (idem. 251-72). Complex Transitive Predicates (P-OD -PCO ) resemble P-PCS
predicates in that the additional direct object does not
generally affect the P-PCS core (e.g. “[consider (the
winner/it/the poison) ideal](+) ”). Hence the ranking
[Head:P-PCO Comp:OD ] (ibidem). (S)ubjects
are ranked as [Head:P Comp:S] (e.g. “[love can
hurt](-) ”, “[the misery ended](+) ”) (idem. 235-43).
Note that [¬] NP complements constitute an exception calling for reverse rankings - consider “[nobody
PHRASES
Pre-head
(Det:(Det|DP)|Subj-Det:NPgen [¬] |Mod:(Neg|AdjP|V))
Head:N
(Det:(Det|DP)|Mod:(Neg|PP|AdvP)) Head:Adj
(Det:(Det|DP)|Mod:(Neg|Adv))
Head:Adv
Mod:(Neg|AdvP|NP)
Head:P
(Subj-Det:NPgen |Mod:(N|Nom)) Head:N
CLAUSES
(Comp:(PCS |S[¬] |OD[¬] |OI[¬] )|A:(AdvP|AdjP|PP)|Mod:Neg)
Head:P
Comp:OD
Head:P-PCO
Head:(N|Nom)
Head:Adj
Head:Adv
Head:P
Head:N
Head:P
Head:P-OD
Post-head
Comp:(AdjP|VP)
Comp:PP
Comp:PP
Comp:(NP|VP)
Comp:(NP|PP)
Comp:(S|OD )
Comp:(OI |OC )
Table 1: Sample Construction Rankings
died](+) ”, “[killed nobody](+) ”, for example. Hence
the rankings [Comp:(OD[¬] |S[¬] ) Head:P] for these
special cases. Adjuncts are generally stronger than
predicators and predicates. The ranking [Comp:AdvP
Head:P] for AdvP Adjuncts, for example, supports propagation (e.g. “[he moved it gently](+) ”), and
polarity conflicts (e.g. “[greeted him insincerely](-) ”)
(idem. 224-5, 575, 669, 779-84).
These and other sample rankings are summarised
in Table 1.
4
Implementation
The proposed model was implemented as a lexical
parsing post-process interpreting the output of a dependency parser1 . We employ a sentiment lexicon
containing manually-compiled atomic core carriers2
expanded semi-automatically using WordNet 2.1, all
tagged with the compositional tags. A morphological
unknown carrier guessing module and a missing dependency link repair module are included. Adhering to
the proposed compositional processes and constituent
rankings at each stage of the analysis, token dependency links and morphosyntactic token tags (e.g. word
class, syntactic role, (pre-/post-)head status) are first
used to construct individual syntactic phrases (NPs,
VPs, AdjPs, AdvPs) and to calculate their internal polarities (phrasal sentiment) through stepwise chunking rules which find the rightmost subconstituent in
a given phrase and expand it leftwards until a phrasal
boundary is hit (see Ex. 2-3). To calculate clausal
and sentential sentiment, the obtained phrasal constituents are then combined (see Ex. 4).
5
Experiments
To estimate the usefulness of a compositional treatment and its impact on standard NLP pipelines, we
employ short headlines for sentential compositionality
and NPs for phrasal compositionality. Since our implementation is fully lexical, its recall is conditioned by
the coverage of the lexicon used. To estimate the (future) impact of larger lexica covering the entire WordNet, the default lexicon (at the time of writing) (DEFAULT LEX) was expanded with sample carriers from
the test data found in WordNet 2.1 (WN ADD LEX).
Polarity agreement between the gold standards and
our output was measured using (i ) all polarities (All
1
2
Connexor Machinese Syntax 3.8 (www.connexor.com)
Kindly provided by Corpora Software
(www.corporasoftware.com)
pol ), and (ii ) non-neutral polarities only (Non-ntr pol ).
To assess the role of sentiment intensity, results using
(i ) cases of Any Strength and (ii ) those marked as
Strong in the gold standards are given. The agreement results are shown in Table 2.
Experiment 1: Headlines. The sentences generated by our system were compared against 1000 news
headlines in the SemEval-2007 Task #14 data set annotated for polarity (six annotators, r .78) ([8]). The
SemEval scores [-100, 100] were collapsed into (-100 ≤
(-) < 0; 0 = (N); 0 < (+) ≤ 100) in the Any Strength
condition, and into (-100 ≤ (-) ≤ -66; 0 = (N); 66 ≤ (+)
≤ 100) for 208 Strong cases. The WN ADD LEX lexicon
contained 97 added carriers.
Experiment 2: NPs. The NPs generated by our
system were compared (lax overlap) against 1541 explicit NPs in the customer review data set of 2108
product feature mentions from five home electronics
products annotated for polarity (two annotators, r unknown) ([3]). The gold standard scores [-3, 3] were
converted into (-3 ≤ (-) < 0; 0 < (+) ≤ 3) in the Any
Strength condition, and into (-3 = (-); 3 = (+)) for 366
Strong cases. The WN ADD LEX lexicon contained 95
added carriers.
Results and Error Analysis
The All pol figures are considerably lower than the
corresponding Non-ntr pol ones due to the incomplete
coverage of the lexica used: a (N) input into the model
leads unavoidably to a (N) output and thus to an error.
Since mining and tagging new carriers is a task beyond
the realms of the model, we focus here on the performance in the Non-ntr pol conditions. Mirroring human judgements of high-intensity cases, the implementation performed noticeably better with strong cases.
More interesting is the small margin between the two
lexica which offers further evidence pro compositionality. The errors from the [WN ADD LEX Non-ntr pol
Any Strength] condition are analysed in Table 3.
Because the model operates in the middle of
the processing pipeline, the errors are classified as
pre-compositional (i.e. erroneous input) or postcompositional (i.e. factors beyond the model). The
performance of the model is promising as most errors (ca. 2/3) occurred earlier in the pipeline. Since
full compositionality can only be achieved with a clean
grammatical analysis, a heavy burden is placed on the
TAGGER and PARSER which together caused ca. 28%
of the errors. Hence erroneous propagation and partial compositionality due to incorrect POS tags and
null dependencies, respectively. Since polarity distinctions between individual word SENSEs (e.g. “[rip
DEFAULT LEX
Cases
All pol
1000
A
63.0
208
A
81.73
1541
P
R
F
72.46
97.79
83.24
366
P
R
F
79.22
98.63
87.87
WN ADD LEX
Non-ntr pol All pol
Headlines
Any Strength
76.27
65.6
Strong
89.95
86.06
NPs
Any Strength
85.45
73.39
82.93
97.79
84.17
83.85
Strong
89.10
80.33
87.70
98.63
88.40
88.55
Table 2:
Agreement:
(R)ecall, and (F)-scores
Non-ntr pol
77.36
91.33
85.87
83.58
84.71
89.51
88.52
89.01
(A)ccuracy, (P)recision,
(a CD)](N) ” vs. “[rip into](-) ”) can have far-reaching
compositional consequences, a sentiment WSD module
could reduce ca. 25% of the errors. Resolving neutral
ANAPHORic and CO-REFerential expressions could increase recall levels further. The errors also include
supraclausal cases NOT yet IMPLemented. However,
even in an ideal situation with a clean input, the model
would fail to solve many cases (ca. 19%) in which further WORLD knowledge is required. There are cases
in which the literal/logical compositional polarity is
modulated by phenomena closer to PRAGMatics than
lexical semantics such as indirect speech acts (cf. logical positivity vs. implied negativity in “[X could be
better](-) ”). Lastly, AMBIGuous cases affording multiple polarity readings are always likely to be present.
ANAPHOR
CO-REF
NOT IMPL
PARSER
SENSE
SPELLING
TAGGER
AMBIG
PRAGM
WORLD
Total
Headlines NPs
Pre-compositional errors
13 (6.05)
4 (1.86)
8 (4.26)
22 (10.23)
13 (6.92)
44 (20.47)
58 (30.85)
46 (21.4)
2 (0.93)
24 (12.77)
32 (14.88)
All
13
4
30
57
104
2
56
266
3.23
0.99
7.44
14.14
25.81
0.5
13.9
66%
Post-compositional errors
35 (18.62)
4 (1.86)
39
21 (9.77)
21
50 (26.6)
27 (12.56) 77
137
188
215
403
9.68
5.21
19.11
34%
100%
Table 3: Error distribution
6
Related Work
The proposed model develops further the lexical devices described in the survey of lexical and discoursal contextual valence shifters in ([7]). In ([6]), negation and change phrases were used in a supervised
learning algorithm analysing sentential polarities of
clinical outcomes. A number of polarity shifters and
syntactic dependencies were included as machine learning features in the phrase-level sentiment analyser
reported in ([10]). Adjectival appraisal groups comprising a head adjective with optional appraisal premodifiers were used in the sentiment classifier described
in ([9]). ([1]) extracted and tagged words with reversal potential expressing a conceptual in-/decrease in
magnitude, intensity or quality.
7
Conclusion
We have shown that sentiment exhibits quasi-compositionality in noticeably many areas, and that it is possible to approach sentiment propagation, polarity reversal, and polarity conflict resolution within different
linguistic constituent types at different grammatical
levels in an analytically and computationally uniform
manner by relying on traditional compositional semantics and deep parsing. The results obtained, which are
encouraging for a lexical system, point towards a crucial dependency on a wide-coverage lexicon, accurate
parsing, and sentiment sense disambiguation in a compositional approach to sentiment analysis.
References
[1] A. Andreevskaia and S. Bergler. Semantic tag extraction using wordnet glosses. In Proceedings of LREC
2006, Genoa, 2006.
[2] D. Dowty, R. Wolf, and S. Peters. Introduction to
Montague Semantics. D. Reidel, Dordrecht, 1981.
[3] M. Hu and B. Liu. Mining and summarizing customer
reviews. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery & Data
Mining (KDD-2004), Seattle, 2004.
[4] R. Huddleston and G. K. Pullum. The Cambridge
Grammar of the English Language. Cambridge University Press, Cambridge, 2002.
[5] S.-M. Kim and E. Hovy. Determining the sentiment
of opinions. In Proceedings of COLING 2004, Geneva,
2004.
[6] Y. Niu, X. Zhu, J. Li, and G. Hirst. Analysis of polarity information in medical text. In Proceedings of
the American Medical Informatics Association 2005
Annual Symposium (AMIA 2005), Washington D.C.,
2005.
[7] L. Polanyi and A. Zaenen. Contextual lexical valence
shifters. In Y. Qu, J. Shanahan, and J. Wiebe, editors, Exploring Attitude and Affect in Text: Theories
and Applications: Papers from the 2004 Spring Symposium, Technical Report SS-04-07. AAAI, 2004.
[8] C. Strapparava and R. Mihalcea. Semeval-2007 task
14: Affective text. In Proceedings of SemEval 2007,
Prague, 2007.
[9] C. Whitelaw, N. Garg, and S. Argamon. Using appraisal taxonomies for sentiment analysis. In Proceedings of the 2005 ACM CIKM International Conference on Information and Knowledge Management,
Bremen, 2005.
[10] T. Wilson, J. Wiebe, and P. Hoffmann. Recognizing
contextual polarity in phrase-level sentiment analysis. In Proceedings of HLT/EMNLP 2005, Vancouver,
2005.
[11] H. Yu and V. Hatzivassiloglou. Towards answering
opinion questions: Separating facts from opinions and
identifying the polarity of opinion sentences. In Proceedings of EMNLP 2003, Sapporo, 2003.